autosnippet
Version:
Extract code patterns into a knowledge base for AI coding assistants
359 lines (358 loc) • 16.1 kB
JavaScript
/**
* EpisodicConsolidator — Episodic → Semantic 固化引擎
*
* Bootstrap 完成后，将 SessionStore (Tier 2) 中的维度分析结果
* 提炼为结构化记忆，固化到 PersistentMemory (Tier 3)。
*
* 固化策略 (规则化，无需额外 AI 调用):
* 1. 从每个维度的 findings 提取 fact 记忆
* 2. 从 Tier Reflections 的 crossDimensionPatterns 提取 insight 记忆
* 3. 从 analysisText 中提取项目级别事实 (正则匹配)
* 4. 使用 PersistentMemory.consolidate() 进行去重和合并
*
* @module EpisodicConsolidator
*/
import Logger from '#infra/logging/Logger.js';
// ──────────────────────────────────────────────────────────────
// 正则: 从分析文本中提取陈述性知识
// ──────────────────────────────────────────────────────────────
/**
* 匹配常见的项目事实陈述模式:
* - "项目使用 XX 模式"
* - "发现 XX 个 YY"
* - "主要语言是 XX"
* - "XX 是唯一的 YY"
* - "XX 采用了 YY"
*/
const FACT_PATTERNS = [
// Chinese
/(?:项目|工程|代码库)(?:使用|采用|基于|遵循)了?\s*([^，。,.\n]{5,60})/g,
/(?:主要|核心|主|主力)\s*(\S+)\s*(?:是|为|使用)\s*([^，。,.\n]{3,40})/g,
/(?:发现|找到|扫描到|识别|共有|包含)\s*了?\s*(\d+)\s*个?\s*([^，。,.\n]{2,30})/g,
/(\S{2,20})\s*是\s*(?:唯一的?|主要的?|核心的?|全局的?)\s*([^，。,.\n]{3,30})/g,
/(?:使用|采用|遵循)了?\s*(\S{1,10})\s*(?:前缀|后缀|命名|约定|规范)/g,
// English
/(?:the\s+)?project\s+(?:uses?|adopts?|relies\s+on|follows?)\s+([^.,\n]{5,60})/gi,
/(?:found|discovered|identified|detected)\s+(\d+)\s+([^.,\n]{3,40})/gi,
/(?:the\s+)?(?:primary|main|core)\s+(\S+)\s+(?:is|are)\s+([^.,\n]{3,40})/gi,
/(?:all|every)\s+([^.,\n]{3,30})\s+(?:use|adopt|follow|implement)\s+([^.,\n]{3,40})/gi,
/(?:there\s+(?:is|are))\s+(\d+)\s+([^.,\n]{3,40})/gi,
];
/**
* 匹配洞察性陈述:
* - "XXX 暗示/表明/说明 YYY"
* - "XXX 与 YYY 耦合/关联"
* - "建议/推荐 XXX"
*/
const INSIGHT_PATTERNS = [
// Chinese
/([^，。,.\n]{5,40})(?:暗示|表明|说明|意味着|揭示)\s*([^，。,.\n]{5,60})/g,
/([^，。,.\n]{3,20})\s*(?:与|和)\s*([^，。,.\n]{3,20})\s*(?:耦合|关联|存在依赖|有关系)/g,
/(?:建议|推荐|应该|需要)\s*([^，。,.\n]{5,60})/g,
// English
/([^.,\n]{5,40})\s+(?:suggests?|indicates?|implies?|reveals?)\s+(?:that\s+)?([^.,\n]{5,60})/gi,
/([^.,\n]{3,20})\s+(?:is|are)\s+(?:tightly\s+)?(?:coupled|linked|related)\s+(?:to|with)\s+([^.,\n]{3,30})/gi,
/(?:recommend|should|consider|suggest)\s+([^.,\n]{5,60})/gi,
];
// ──────────────────────────────────────────────────────────────
// EpisodicConsolidator 类
// ──────────────────────────────────────────────────────────────
export class EpisodicConsolidator {
#semanticMemory;
#logger;
constructor(semanticMemory, { logger } = {}) {
this.#semanticMemory = semanticMemory;
this.#logger = logger || Logger.getInstance();
}
/**
* 执行固化: SessionStore → PersistentMemory
*
* @param [opts.bootstrapSession] Bootstrap session ID
* @param [opts.clearPrevious=false] 是否先清除旧的 bootstrap 记忆
* @returns }
*/
consolidate(sessionStore, { bootstrapSession, clearPrevious = false, } = {}) {
const t0 = Date.now();
// 可选: 清除旧的 bootstrap 记忆 (全量重跑场景)
if (clearPrevious) {
const cleared = this.#semanticMemory.clearBootstrapMemories();
this.#logger.info(`[Consolidator] Cleared ${cleared} previous bootstrap memories`);
}
// 1. 先执行维护 (过期清理)
this.#semanticMemory.compact();
// 2. 从 findings 提取 fact 记忆
const findingMemories = this.#extractFromFindings(sessionStore);
// 3. 从 Tier Reflections 提取 insight 记忆
const insightMemories = this.#extractFromReflections(sessionStore);
// 4. 从 analysisText 提取文本中的事实
const textFactMemories = this.#extractFromAnalysisText(sessionStore);
// 5. 合并所有候选, 使用 consolidate 去重
const allCandidates = [...findingMemories, ...insightMemories, ...textFactMemories];
// ── 结构化统计日志 ──
const dimStats = this.#computeDimStats(allCandidates);
const importanceDist = this.#computeImportanceDistribution(allCandidates);
const entityCount = allCandidates.reduce((sum, c) => sum + (c.relatedEntities?.length || 0), 0);
this.#logger.info(`[Consolidator] Extracted ${allCandidates.length} candidate memories: ` +
`${findingMemories.length} findings, ${insightMemories.length} insights, ` +
`${textFactMemories.length} text facts`);
this.#logger.info(`[Consolidator] Per-dimension: ${dimStats.map((d) => `${d.dim}=${d.count}`).join(', ')}`);
this.#logger.info(`[Consolidator] Importance distribution: ${importanceDist} | Entities extracted: ${entityCount}`);
const result = this.#semanticMemory.consolidate(allCandidates, { bootstrapSession });
const durationMs = Date.now() - t0;
this.#logger.info(`[Consolidator] Consolidation complete in ${durationMs}ms: ` +
`+${result.added} ADD, ~${result.updated} UPDATE, ⊕${result.merged} MERGE, ` +
`=${result.skipped} SKIP`);
return {
findings: { extracted: findingMemories.length },
insights: { extracted: insightMemories.length },
textFacts: { extracted: textFactMemories.length },
total: result,
durationMs,
perDimension: Object.fromEntries(dimStats.map((d) => [d.dim, d.count])),
importanceDistribution: this.#importanceHistogram(allCandidates),
entityCount,
};
}
// ─── 提取器 ───────────────────────────────────────────
/**
* 从维度 findings 提取 fact 记忆
*
* 每个 finding 映射为一条 fact，importance 直接继承。
*/
#extractFromFindings(sessionStore) {
const memories = [];
const completedDims = sessionStore.getCompletedDimensions();
for (const dimId of completedDims) {
const report = sessionStore.getDimensionReport(dimId);
if (!report?.findings) {
continue;
}
for (const f of report.findings) {
// 跳过低重要性的发现
if ((f.importance || 5) < 4) {
continue;
}
// 跳过过短的发现
const content = typeof f === 'string' ? f : f.finding || '';
if (content.length < 10) {
continue;
}
// 提取关联实体 (从 evidence 中提取文件名/类名)
const entities = this.#extractEntities(content, f.evidence);
memories.push({
type: 'fact',
content: content.substring(0, 500),
source: 'bootstrap',
importance: typeof f === 'string' ? 5 : f.importance || 5,
sourceDimension: dimId,
sourceEvidence: typeof f === 'string' ? '' : f.evidence || '',
relatedEntities: entities,
tags: [dimId],
});
}
}
return memories;
}
/**
* 从 Tier Reflections 提取 insight 记忆
*
* crossDimensionPatterns → insight (跨维度观察)
* suggestionsForNextTier → insight (分析建议)
* topFindings 中重要性 ≥ 7 的 → fact (高优先级重复确认)
*/
#extractFromReflections(sessionStore) {
const memories = [];
const json = sessionStore.toJSON();
const reflections = json.tierReflections || [];
for (const ref of reflections) {
// 跨维度模式 → insight
for (const pattern of ref.crossDimensionPatterns || []) {
if (pattern.length < 10) {
continue;
}
memories.push({
type: 'insight',
content: pattern.substring(0, 500),
source: 'bootstrap',
importance: 7, // 跨维度发现通常较重要
sourceDimension: `tier-${ref.tierIndex + 1}-reflection`,
relatedEntities: this.#extractEntities(pattern),
tags: ref.completedDimensions || [],
});
}
// 建议 → insight (较低优先级)
for (const suggestion of ref.suggestionsForNextTier || []) {
if (suggestion.length < 10) {
continue;
}
memories.push({
type: 'insight',
content: suggestion.substring(0, 500),
source: 'bootstrap',
importance: 5,
sourceDimension: `tier-${ref.tierIndex + 1}-reflection`,
tags: ['suggestion'],
});
}
// 高重要性 topFindings → fact (≥ 7 分的重要发现)
for (const f of ref.topFindings || []) {
if ((f.importance || 5) < 7) {
continue;
}
const content = typeof f === 'string' ? f : f.finding || '';
if (content.length < 10) {
continue;
}
memories.push({
type: 'fact',
content: content.substring(0, 500),
source: 'bootstrap',
importance: f.importance || 7,
sourceDimension: f.dimId || `tier-${ref.tierIndex + 1}`,
sourceEvidence: f.evidence || '',
relatedEntities: this.#extractEntities(content),
tags: [f.dimId, 'tier-reflection'].filter(Boolean),
});
}
}
return memories;
}
/**
* 从分析文本中正则提取项目级事实和洞察
*
* 仅提取高置信度的简短陈述 (≤100 字), 避免噪音。
*/
#extractFromAnalysisText(sessionStore) {
const memories = [];
const seen = new Set(); // 去重
const completedDims = sessionStore.getCompletedDimensions();
for (const dimId of completedDims) {
const report = sessionStore.getDimensionReport(dimId);
if (!report?.analysisText) {
continue;
}
const text = report.analysisText;
// 提取事实
for (const pattern of FACT_PATTERNS) {
// 重置 lastIndex (全局正则)
pattern.lastIndex = 0;
let match;
let matchCount = 0;
while ((match = pattern.exec(text)) !== null && matchCount < 5) {
const fullMatch = match[0].trim();
if (fullMatch.length < 10 || fullMatch.length > 120) {
continue;
}
if (seen.has(fullMatch)) {
continue;
}
seen.add(fullMatch);
matchCount++;
memories.push({
type: 'fact',
content: fullMatch,
source: 'bootstrap',
importance: 4, // 正则提取的置信度偏低
sourceDimension: dimId,
relatedEntities: this.#extractEntities(fullMatch),
tags: [dimId, 'text-extracted'],
});
}
}
// 提取洞察
for (const pattern of INSIGHT_PATTERNS) {
pattern.lastIndex = 0;
let match;
let matchCount = 0;
while ((match = pattern.exec(text)) !== null && matchCount < 3) {
const fullMatch = match[0].trim();
if (fullMatch.length < 10 || fullMatch.length > 120) {
continue;
}
if (seen.has(fullMatch)) {
continue;
}
seen.add(fullMatch);
matchCount++;
memories.push({
type: 'insight',
content: fullMatch,
source: 'bootstrap',
importance: 4,
sourceDimension: dimId,
relatedEntities: this.#extractEntities(fullMatch),
tags: [dimId, 'text-extracted'],
});
}
}
}
return memories;
}
// ─── 辅助方法 ─────────────────────────────────────────
/** 按维度聚合候选数量 */
#computeDimStats(candidates) {
const counts = new Map();
for (const c of candidates) {
const dim = c.sourceDimension || 'unknown';
counts.set(dim, (counts.get(dim) || 0) + 1);
}
return [...counts.entries()]
.map(([dim, count]) => ({ dim, count }))
.sort((a, b) => b.count - a.count);
}
/** 生成重要性分布字符串: "[1-3]=N [4-6]=N [7-10]=N" */
#computeImportanceDistribution(candidates) {
let low = 0;
let mid = 0;
let high = 0;
for (const c of candidates) {
const imp = c.importance || 5;
if (imp <= 3) {
low++;
}
else if (imp <= 6) {
mid++;
}
else {
high++;
}
}
return `[1-3]=${low} [4-6]=${mid} [7-10]=${high}`;
}
/** 构建重要性直方图对象 (供返回值使用) */
#importanceHistogram(candidates) {
const hist = {};
for (const c of candidates) {
const imp = c.importance || 5;
hist[imp] = (hist[imp] || 0) + 1;
}
return hist;
}
/**
* 从文本中提取实体名 (类名/文件名/模块名)
*
* 简单规则:
* - 大驼峰式: BDNetworkManager, UIViewController
* - 文件路径: Classes/Network/BDRequest.m
* - 冒号分隔的 evidence: "BDRequest.m:42"
*/
#extractEntities(text, evidence = undefined) {
const entities = new Set();
// 大驼峰类名 (至少 2 个大写字母)
const classNames = (text || '').match(/\b[A-Z][a-zA-Z]*[A-Z][a-zA-Z]*\b/g) || [];
for (const name of classNames) {
if (name.length >= 4 && name.length <= 40) {
entities.add(name);
}
}
// 从 evidence 提取文件名
if (evidence) {
const fileName = evidence.split(':')[0].split('/').pop();
if (fileName && fileName.length >= 3) {
entities.add(fileName);
}
}
return [...entities].slice(0, 5); // 最多 5 个实体
}
}
export default EpisodicConsolidator;